PyTorch-GAN-Experiments-Env

PyTorch GAN playground. Features:

1. yaml config file
2. Easy interface for Tensorboard
3. Generic system - just plug and play

Implemented very simple versions of GAN and ACGAN

Project Structure

.
├── config                  # Config files (.yml, .json) using with hb-config
├── gan                     # Generative Adversarial Networks architecture 
    ├── __init__.py             # train, evaluate, predict logic
    ├── module.py               # Discriminator, Generator module
    └── utils.py                # Save and Load Model, TensorBoard
├── acgan                     # ACGAN architecture
    ├── __init__.py             # train,logic
    ├── module.py               # Discriminator, Generator module
    └── utils.py                # Save and Load Model, TensorBoard
├── data_loader.py          # make_data_loader (using DataLoader)
├── main.py                 
├── model.py                # define Model Spec
└── utils.py                # utils
└── plots.py                # generic functions to plot you results from a pickle file
└── sampler.py              # different sampling methods:multi-modal uniform,muli-modal Gaussian,truncated normal...

Reference : hb-config

• Manage experiments like tf.Estimator

Create a new model

In order to implement your own model just create a new branch, copy the gan directory and edit it as you want.

Data loader

Pytorch supports many datasets which are implemented to work with the data loaders, so if you want to work on different dataset just add it in the data_loader.py file and repsectively in the config file. Supported data sets

Config

Can control all Experimental environment.

example: config.yml

data:
  path: "data/"

model:
  z_dim: 100     # Random noise dimension coming into generator, per output vector
  real_dim: 784

  g_h1: 256
  g_h2: 512
  g_h3: 1024

  d_h1: 1024
  d_h2: 512
  d_h3: 256

  dropout: 0.3

train:
  model_dir: "logs/gan"
  batch_size: 64
  train_steps: 50000

  d_learning_rate: 0.0002  # 2e-4
  g_learning_rate: 0.0002
  optim_betas:
    - 0.9
    - 0.999

  save_checkpoints_steps: 1000
  verbose_step_count: 100

predict:
  batch_size: 64

slack:
  webhook_url: ""  # after training notify you using slack-webhook

Usage

Install requirements.

pip install -r requirements.txt

Then, start training

python main.py --mode train

After training, generate images

python main.py --mode predict

Some code adapted from Dongjun Lee